fix bugs when token_classify & classify run concurrently (#36614)

Signed-off-by: augusto.yjh <augusto.yjh@antgroup.com>

vllm/model_executor/layers/pooler/tokwise/methods.py

@@ -47,10 +47,13 @@ class AllPool(TokenPoolingMethod):
         pooling_metadata: PoolingMetadata,
     ) -> list[TokenPoolingMethodOutputItem]:
         pooling_cursor = pooling_metadata.get_pooling_cursor()
-        hidden_states_all = hidden_states.split(
-            pooling_cursor.num_scheduled_tokens_cpu.tolist()
+        hidden_states_lst = [
+            hidden_states[first : last + 1]
+            for first, last in zip(
+                pooling_cursor.first_token_indices_gpu.tolist(),
+                pooling_cursor.last_token_indices_gpu.tolist(),
             )
-        hidden_states_lst = [hidden_states_all[i] for i in pooling_cursor.index]
+        ]
 
         if not self.enable_chunked_prefill:
             return hidden_states_lst